Development of a Methodological Framework for Hyperspectral Estimation of Water Quality Parameters in Diverse Inland Waters

Water quality is a fundamental indicator of ecosystem health, with significant influence on biodiversity conservation, public health, and sustainable management of natural resources. Parameters such as chlorophyll-a (Chl-a) and total suspended solids (TSS) are crucial for monitoring changes in aquatic ecosystems and identifying critical issues like eutrophication and harmful algal blooms (HABs). These processes are exacerbated by climate change, which intensifies extreme events, alters hydrological cycles, and increases nutrient runoff, with profound impacts on water bodies globally. In this context, the present study proposes the development of specialized algorithms to estimate water quality parameters in various types of inland waters, such as clear, turbid, and chlorophyll-rich waters. The implemented methodology establishes an advanced processing framework to predict water quality parameters in environments with diverse optical properties. To classify optical water types (OWTs), a Random Forest classifier was trained using the GLORIA database (a globally representative hyperspectral in situ dataset), preprocessed with unsupervised clustering algorithms and reflectance’s simulated for the Italian Space Agency (ASI)’s hyperspectral PRISMA sensor. Subsequently, regression models based on Random Forest, Support Vector Machine (SVM), and XGBoost were trained to predict key parameters such as Chl-a concentration and TSS. Independent variables for these models were selected through a sensitivity analysis of the simulated PRISMA sensor channels. Results were achieved over a wide selection of lakes in Italy (e.g. Trasimeno Lake) and around the world, with more than 2,000 in-situ samples. Among the key outcomes that will be presented, it is worth highlighting that: OWT Classification: spectral clustering succeeded in capturing a greater diversity in optical behavior and also maintains reduced variance for each of the selected classes. Although an unsupervised approach allows identification of patterns inherent to the dataset, exploring supervised classifications based on the spectral similarity of modeled OWTs offers greater control in training a classifier. Model Performance Chl-a: Cross-validation evaluations indicate that the random forest model offers the best performance with R2 higher than 0.6. However, in waters with Chl-a concentrations below 10 mg/m³, overestimations of around 5 mg/m³ were observed, while for concentrations above 200 mg/m³, the model tends to underestimate by approximately 90 mg/m³. Sensitivity Analysis: The analysis revealed how well individual PRISMA channels correlated with Chl-a concentrations across the nine OWTs. Based on the coefficient of determination (R²), certain PRISMA channels are more sensitive to Chl-a in specific OWTs. This insight highlights the importance of spectral resolution and the need to tailor models based on water type for accurate water quality assessments. This approach therefore appears promising in addressing the challenge of accurately estimating water quality parameters in different aquatic environments, tailoring the algorithms to the specific characteristics of each water type. As a result, it enhances the precision of parameter estimation and optimizes the use of hyperspectral data for local applications. This research is performed in the framework of the SatellOmic project, funded by the Italian Space Agency (ASI), Agreement n. 2023-36-HH.0, as part of the ASI’s program “Innovation for Downstream Preparation for Science” (I4DP_SCIENCE).